Research map / Continual learning / Online and replay-based
Active Learning for Continual Learning: Keeping the Past Alive in the Present
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AccuACL, an active continual learning query strategy that uses a Fisher information embedding over a rehearsal memory buffer and the current unlabeled pool to select samples that balance forgetting prevention and new-task learning.
Problem
Conventional active learning strategies query examples only informative for the new task, causing catastrophic forgetting in active continual learning (ACL) where labels are costly and past data is inaccessible.
Approach
Defines accumulated informativeness via the target Fisher information decomposed into past (memory buffer) and new (unlabeled pool) parts with a theoretically derived balancing weight, approximates Fisher information with a per-example diagonal embedding from the last linear layer, and greedily selects samples by a Jensen-Shannon distribution score followed by a magnitude score.
Key results
- AccuACL improves average accuracy and forgetting over AL baselines by 23.8% and 17.0% on average across CL methods and datasets.
- On SplitCIFAR100 (M=1000) with DER++, AccuACL reaches A10 of 36.3 vs 35.9 for Uniform and forgetting of 15.0 vs 21.5 for Uniform.
- BAIT requires 462.1 times more querying time and 85.5 times more memory than AccuACL on SplitCIFAR10 and runs out of memory on SplitCIFAR100 and SplitTinyImageNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AccuACL instead of conventional active-learning baselines | DER++ · SplitCIFAR100 (M=500) | average accuracy: 27.4 -> 30.0 | Helped |
| use AccuACL instead of conventional active-learning baselines | DER++ · SplitCIFAR100 (M=500) | forgetting: 38.7 -> 29.3 | Helped |
| use AccuACL instead of conventional active-learning baselines | DER++ · SplitCIFAR100 (M=500) | learning accuracy: 65.9 -> 56.6 | Hurt |
| estimate target Fisher information using all previously seen task data instead of a memory buffer (ablation) | AccuACL on DER++ · SplitCIFAR100 (M=1000) | average accuracy: 36.3±0.4 -> 33.9±0.9 | Hurt |
| estimate target Fisher information using all previously seen task data instead of a memory buffer (ablation) | AccuACL on DER++ · SplitCIFAR100 (M=1000) | forgetting: 15.0±0.5 -> 14.2±0.8 | Helped |
| increase the past-task weight λ (ablation) | AccuACL on ER · SplitCIFAR10, task 4 | forgetting | Helped |
| set λ to its theoretical value rather than another tested value (ablation) | AccuACL on ER · SplitCIFAR10, task 4 | accuracy | Helped |
Surprising findings
- Conventional AL strategies often underperform even uniform random querying in ACL due to bias toward new tasks.
- Adding the memory buffer as an initial selection to kCenter and BADGE degraded their performance.
Limitations
- Relies on a rehearsal memory buffer; rehearsal-free variant not validated.
- Not evaluated under realistic CL settings such as noisy labels, blurry tasks, or imbalanced data.
- Only class-incremental image classification benchmarks with ResNet18 are evaluated.
Experimental setup
Datasets: SplitCIFAR10 SplitCIFAR100 SplitTinyImageNet CIFAR10 CIFAR100 TinyImageNet
Benchmarks: SplitCIFAR10 SplitCIFAR100 SplitTinyImageNet
Metrics: average accuracy forgetting learning accuracy
Baselines: Uniform Entropy LeastConfidence kCenterGreedy BADGE BAIT
Builds on: BAIT Fisher information-based active learning ER DER++ GSS ER-ACE
Closest papers by meaning
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.88)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · similarity 0.88)
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.87)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.87)
- Advancing Prompt-Based Methods for Replay-Independent General Continual Learning (ICLR 2025 · similarity 0.87)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.